Multi-Time Resolution Ensemble LSTMs for Enhanced Feature Extraction in High-Rate Time Series

Barzegar, Vahid; Laflamme, Simon; Hu, Chao; Dodson, Jacob

doi:10.3390/s21061954

Cited by 11 publications

(10 citation statements)

References 38 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Remark that, here, only some levels of constraints on the dynamics learned by the LSTM cells were enforced through the multi-rate sampler. The incorporation of additional physical constraints would require additional physical knowledge, for example, knowledge of modal properties [45], which was taken as not accessible for the drop tower experiment. Fig.…”

Section: Numerical Resultsmentioning

confidence: 99%

Ensemble of recurrent neural networks with long short-term memory cells for high-rate structural health monitoring

Barzegar

Laflamme

et al. 2022

Mechanical Systems and Signal Processing

Self Cite

View full text Add to dashboard Cite

The deployment of systems experiencing high-rate dynamic events, such as hypersonic vehicles, advanced weaponry, and active blast mitigation systems, require high-rate structural health monitoring (HRSHM) capabilities in the sub-millisecond realm to ensure continuous operations and safety. However, the development of high-rate feedback systems is a complex task because these dynamic systems are uniquely characterized by (1) large uncertainties in their external loads, (2) high levels of non-stationarity and heavy disturbance, and (3) unmodeled dynamics from changes in system configuration. In this paper, we present a deep learning algorithm specifically engineered for HRSHM applications. It consists of an ensemble of recurrent neural networks (RNNs) constructed with long short-term memory (LSTM) cells with transfer learning capabilities to cope with the highly limited availability of training data as it is typical for high-rate systems. The use of an ensemble of RNNs empowers multi-rate sampling capability to capture multi-temporal features of the time series, thus enabling modeling of non-stationarities. Also, because the RNNs use short-sequence LSTMs and are arranged in parallel, the computation time is substantially reduced to the sub-millisecond range. The performance of the algorithm is investigated on experimental high-rate dynamic data produced from accelerated drop tower tests and benchmarked against results from a prior investigation using a purely on-the-edge algorithm termed variable input space observer (VIO) and its hybrid architecture (hybrid VIO) that incorporated physical knowledge and is considered as an upper bound on the algorithm performance. Numerical results show that the ensemble of RNNs, composed of five RNNs, significantly outperformed the VIO, with performance close to that of the hybrid VIO when it comes to the estimation error metrics, with an average computation time of 25 μs per prediction step. Yet, the ensemble of RNNs exhibits chattering for low-amplitude excitations, likely attributable to the behavior of RNNs sampling at small time delays. An examination of the RNN weights and hidden states confirms that the algorithm can capture multi-temporal features, and an investigation with respect to noise in the training dataset showed that the algorithm is robust for signal-to-noise ratios up to 20 dB.

show abstract

Section: Numerical Resultsmentioning

confidence: 99%

Ensemble of recurrent neural networks with long short-term memory cells for high-rate structural health monitoring

Barzegar

Laflamme

et al. 2022

Mechanical Systems and Signal Processing

Self Cite

View full text Add to dashboard Cite

show abstract

“…Time series are explored in several types of applications and but not the subject of study until today. Most of these applications are focused on forecasting [32,33] and feature extraction [5] approaches, as can be seen in the following studies.…”

Section: Related Workmentioning

confidence: 99%

“…Its operation considers that the next value to be obtained will be equal to the last observed value. Equation (5) shows how these values are obtained:…”

Section: Prediction Techniquesmentioning

confidence: 99%

“…Traditionally, stochastic processes are used to model time series behavior with great success [3,4]. In addition, machine learning-based approaches are also employed to perform the identification of complex behaviors of nonlinear patterns, optimization of unconventional functions, and even establishing connections with long dependencies through recurrent neural networks [5,6]. These patterns can be verified in different areas, such as climatic data [7], sales [8], medical diagnosis [9][10][11], security [1,12], and even the change in share values on the stock exchange [13].…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Time Series Segmentation Based on Stationarity Analysis to Improve New Samples Prediction

Silva

Zarpelão

Cano

et al. 2021

Sensors

View full text Add to dashboard Cite

A wide range of applications based on sequential data, named time series, have become increasingly popular in recent years, mainly those based on the Internet of Things (IoT). Several different machine learning algorithms exploit the patterns extracted from sequential data to support multiple tasks. However, this data can suffer from unreliable readings that can lead to low accuracy models due to the low-quality training sets available. Detecting the change point between high representative segments is an important ally to find and thread biased subsequences. By constructing a framework based on the Augmented Dickey-Fuller (ADF) test for data stationarity, two proposals to automatically segment subsequences in a time series were developed. The former proposal, called Change Detector segmentation, relies on change detection methods of data stream mining. The latter, called ADF-based segmentation, is constructed on a new change detector derived from the ADF test only. Experiments over real-file IoT databases and benchmarks showed the improvement provided by our proposals for prediction tasks with traditional Autoregressive integrated moving average (ARIMA) and Deep Learning (Long short-term memory and Temporal Convolutional Networks) methods. Results obtained by the Long short-term memory predictive model reduced the relative prediction error from 1 to 0.67, compared to time series without segmentation.

show abstract

“…A particularity of the technique is that it leveraged an input space that varied depending on the time series' local dynamics. An improvement of the algorithm was proposed in Barzegar et al 7 that consisted of an ensemble of recurrent neural networks (RNNs) with long short-term memory (LSTM) cells arranged in parallel, where each RNN was constructed with a different input space in order to provide multi-time resolution capabilities. The algorithm showed great predictive capabilities with an average computation time of 25 µs.…”

Section: Introductionmentioning

confidence: 99%

Long short-term memory networks with attention learning for high-rate structural health monitoring

Barzegar

Laflamme

et al. 2021

Sensors and Smart Structures Technologies for Civil, Mechanical, and Aerospace Systems 2021

Self Cite

View full text Add to dashboard Cite

High-rate dynamic systems undergo events of amplitudes greater than 100 g s in a span of less than 100 ms. The unique characteristics of high-rate dynamic systems include 1) large uncertainties in the external loads, 2) high levels of non-stationarity and heavy disturbances, and 3) unmolded dynamics generated from changes in the system configurations. This paper presents a deep learning algorithm consisting of an ensemble of long short-term memory (LSTM) cells used to conduct high-rate state estimation. The ensemble of LSTMs receives and transforms the signal into inputs of different time resolutions. Each input vector correlates to an LSTM cell which predicts the signal in real-time and produces feature vectors. The feature vectors are then processed through an attention layer and dense layer to predict the physical features of the system. Here, we study the temporal evolution of the attention layer weights to conduct state estimation, while the LSTM cells are attempting to conduct measurement predictions. We study the performance of the algorithm on experimental data generated by DROPBEAR, a dedicated testbed for high-rate structural health monitoring research. State estimation consists of estimating, in real-time, the location of a cart that moves along a beam. Results show that the attention layer weights can be used to estimate the cart location but that the beam requires impact excitations to accelerate the convergence of the algorithm.

show abstract

Multi-Time Resolution Ensemble LSTMs for Enhanced Feature Extraction in High-Rate Time Series

Cited by 11 publications

References 38 publications

Ensemble of recurrent neural networks with long short-term memory cells for high-rate structural health monitoring

Ensemble of recurrent neural networks with long short-term memory cells for high-rate structural health monitoring

Time Series Segmentation Based on Stationarity Analysis to Improve New Samples Prediction

Long short-term memory networks with attention learning for high-rate structural health monitoring

Contact Info

Product

Resources

About